Yes, it's a pasta machine.
But surprisingly, it demonstrates how an artificial neural network for image recognition actually works.
-
In the video you can see that a single image of a dog can be split into four — and in each of the smaller images, the dog is still recognizable.
The reason is that the dog's general outlines are preserved in every one of them, so despite the lower quality, the original subject remains visible.
-
A convolutional neural network — a type of electronic neural network — operates on a similar principle to understand images.
When an image is fed into it, the network applies a series of filters, each designed to detect key contours within the image.
For example, one filter focuses on finding the boundary lines between the subject and the background behind it. Another detects differences between dark and light regions, and so on.
The core idea is that by isolating the object's main outlines from the rest of the image's detail, the network can understand what the object looks like and what defines it.
-
If you think about it, you can easily draw a dog with a pen, a blank sheet of paper, and just a few simple lines.
You could also spend hours drawing every single hair in its coat, adding a detailed background and colors.
Artificial intelligence focuses on separating the key contours from the overall image — and that's how it learns what a dog looks like.
-
In the process of generating an image with artificial intelligence, the workflow runs in reverse.
First, a meaningless image filled with noise is created.
Next, the AI engine sketches the general outlines of the requested subject into it.
In the following stages, detail and color are added, and the quality improves again and again until the final result is produced.